The Medical Segmentation Decathlon

Nature Communications, Journal Year: 2022, Volume and Issue: 13(1)

Published: July 15, 2022

International challenges have become the de facto standard for comparative assessment of image analysis algorithms given a specific task. Segmentation is so far most widely investigated medical processing task, but various segmentation typically been organized in isolation, such that algorithm development was driven by need to tackle single clinical problem. We hypothesized method capable performing well on multiple tasks will generalize previously unseen task and potentially outperform custom-designed solution. To investigate hypothesis, we Medical Decathlon (MSD) - biomedical challenge, which compete multitude both modalities. The underlying data set designed explore axis difficulties encountered when dealing with images, as small sets, unbalanced labels, multi-site objects. MSD challenge confirmed consistent good performance preserved their average different tasks. Moreover, monitoring winner two years, found this continued generalizing wide range other problems, further confirming our hypothesis. Three main conclusions can be drawn from study: (1) state-of-the-art are mature, accurate, retrained tasks; (2) algorithmic across strong surrogate generalizability; (3) training accurate AI models now commoditized non experts.

Language: Английский

---

TorchIO: A Python library for efficient loading, preprocessing, augmentation and patch-based sampling of medical images in deep learning

Computer Methods and Programs in Biomedicine, Journal Year: 2021, Volume and Issue: 208, P. 106236 - 106236

Published: June 17, 2021

Processing of medical images such as MRI or CT presents different challenges compared to RGB typically used in computer vision. These include a lack labels for large datasets, high computational costs, and the need metadata describe physical properties voxels. Data augmentation is artificially increase size training datasets. Training with image subvolumes patches decreases power. Spatial needs be carefully taken into account order ensure correct alignment orientation volumes.We present TorchIO, an open-source Python library enable efficient loading, preprocessing, patch-based sampling deep learning. TorchIO follows style PyTorch integrates standard processing libraries efficiently process during neural networks. transforms can easily composed, reproduced, traced extended. Most inverted, making suitable test-time estimation aleatoric uncertainty context segmentation. We provide multiple generic preprocessing operations well simulation MRI-specific artifacts.Source code, comprehensive tutorials extensive documentation found at http://torchio.rtfd.io/. The package installed from Package Index (PyPI) running pip install torchio. It includes command-line interface which allows users apply files without using Python. Additionally, we graphical user within extension 3D Slicer visualize effects transforms.TorchIO was developed help researchers standardize pipelines allow them focus on learning experiments. encourages good open-science practices, it supports experiment reproducibility version-controlled so that software cited precisely. Due its modularity, compatible other frameworks images.

Language: Английский

---

Artificial intelligence and machine learning for medical imaging: A technology review

Physica Medica, Journal Year: 2021, Volume and Issue: 83, P. 242 - 256

Published: March 1, 2021

Language: Английский

---

Metrics to evaluate the performance of auto-segmentation for radiation treatment planning: A critical review

Radiotherapy and Oncology, Journal Year: 2021, Volume and Issue: 160, P. 185 - 191

Published: May 11, 2021

Language: Английский

---

Head and neck tumor segmentation in PET/CT: The HECKTOR challenge

Medical Image Analysis, Journal Year: 2021, Volume and Issue: 77, P. 102336 - 102336

Published: Dec. 25, 2021

This paper relates the post-analysis of first edition HEad and neCK TumOR (HECKTOR) challenge. challenge was held as a satellite event 23rd International Conference on Medical Image Computing Computer-Assisted Intervention (MICCAI) 2020, its kind focusing lesion segmentation in combined FDG-PET CT image modalities. The challenge's task is automatic Gross Tumor Volume (GTV) Head Neck (H&N) oropharyngeal primary tumors FDG-PET/CT images. To this end, participants were given training set 201 cases from four different centers their methods tested held-out 53 fifth center. ranked according to Dice Score Coefficient (DSC) averaged across all test cases. An additional inter-observer agreement study organized assess difficulty human perspective. 64 teams registered challenge, among which 10 provided detailing approach. best method obtained an average DSC 0.7591, showing large improvement over our proposed baseline agreement, associated with DSCs 0.6610 0.61, respectively. proved successfully leverage wealth metabolic structural properties PET modalities, significantly outperforming level, semi-automatic thresholding based images well other single modality-based methods. promising performance one step forward towards large-scale radiomics studies H&N cancer, obviating need for error-prone time-consuming manual delineation GTVs.

Language: Английский

---

AFTer-UNet: Axial Fusion Transformer UNet for Medical Image Segmentation

2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Journal Year: 2022, Volume and Issue: unknown

Published: Jan. 1, 2022

Recent advances in transformer-based models have drawn attention to exploring these techniques medical image segmentation, especially conjunction with the UNet model (or its variants), which has shown great success under both 2D and 3D settings. Current based methods either directly replace convolutional layers pure transformers or consider a transformer as an additional intermediate encoder between decoder of U-Net. However, approaches only encoding within one single slice do not utilize axial-axis information naturally provided by volume. In setting, convolution on volumetric data consume large GPU memory. One downsample use cropped local patches reduce memory usage, limits performance. this paper, we propose Axial Fusion Transformer (AFTer-UNet), takes advantages layers' capability extracting detailed features transformers' strength long sequence modeling. It considers intra-slice inter-slice long-range cues guide segmentation. Meanwhile, it fewer parameters less train than previous models. Extensive experiments three multi-organ segmentation datasets demonstrate that our method outperforms current state-of-the-art methods.

Language: Английский

---

Fast and Low-GPU-memory abdomen CT organ segmentation: The FLARE challenge

Medical Image Analysis, Journal Year: 2022, Volume and Issue: 82, P. 102616 - 102616

Published: Sept. 13, 2022

Language: Английский

---

Automated detection and segmentation of non-small cell lung cancer computed tomography images

Nature Communications, Journal Year: 2022, Volume and Issue: 13(1)

Published: June 14, 2022

Abstract Detection and segmentation of abnormalities on medical images is highly important for patient management including diagnosis, radiotherapy, response evaluation, as well quantitative image research. We present a fully automated pipeline the detection volumetric non-small cell lung cancer (NSCLC) developed validated 1328 thoracic CT scans from 8 institutions. Along with performance detailed by slice thickness, tumor size, interpretation difficulty, location, we report an in-silico prospective clinical trial, where show that proposed method faster more reproducible compared to experts. Moreover, demonstrate average, radiologists & radiation oncologists preferred automatic segmentations in 56% cases. Additionally, evaluate prognostic power contours applying RECIST criteria measuring volumes. Segmentations our stratified patients into low high survival groups higher significance those methods based manual contours.

Language: Английский

---

Deep learning empowered volume delineation of whole-body organs-at-risk for accelerated radiotherapy

Nature Communications, Journal Year: 2022, Volume and Issue: 13(1)

Published: Nov. 2, 2022

In radiotherapy for cancer patients, an indispensable process is to delineate organs-at-risk (OARs) and tumors. However, it the most time-consuming step as manual delineation always required from radiation oncologists. Herein, we propose a lightweight deep learning framework treatment planning (RTP), named RTP-Net, promote automatic, rapid, precise initialization of whole-body OARs Briefly, implements cascade coarse-to-fine segmentation, with adaptive module both small large organs, attention mechanisms organs boundaries. Our experiments show three merits: 1) Extensively evaluates on 67 tasks large-scale dataset 28,581 cases; 2) Demonstrates comparable or superior accuracy average Dice 0.95; 3) Achieves near real-time in <2 s. This could be utilized accelerate contouring All-in-One scheme, thus greatly shorten turnaround time patients.

Language: Английский

---

A multi-centre polyp detection and segmentation dataset for generalisability assessment

Scientific Data, Journal Year: 2023, Volume and Issue: 10(1)

Published: Feb. 6, 2023

Abstract Polyps in the colon are widely known cancer precursors identified by colonoscopy. Whilst most polyps benign, polyp’s number, size and surface structure linked to risk of cancer. Several methods have been developed automate polyp detection segmentation. However, main issue is that they not tested rigorously on a large multicentre purpose-built dataset, one reason being lack comprehensive public dataset. As result, may generalise different population datasets. To this extent, we curated dataset from six unique centres incorporating more than 300 patients. The includes both single frame sequence data with 3762 annotated labels precise delineation boundaries verified senior gastroenterologists. our knowledge, pixel-level segmentation (referred as PolypGen ) team computational scientists expert paper provides insight into construction annotation strategies, quality assurance, technical validation.

Language: Английский

---